Mistral · Apache-2.0 · HF: mistralai/Mixtral-8x22B-Instruct-v0.1
Mixtral 8x22B Instruct — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$2.00 in / $6.00 out per 1M
Parameters
141B (39B act.)
Attention
GQA (8 KV / 48 Q)
Layers × Hidden
56 × 6144
Max Native Context
64k tokens
MoE Experts
2 / 8
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 267.89 GiB | 271.03 GiB | 276.28 GiB | Lossless Reference |
| Q8_0 | 8.50 | 142.31 GiB | 145.46 GiB | 150.71 GiB | Near-Lossless |
| Q6_K | 6.57 | 110.00 GiB | 113.14 GiB | 118.39 GiB | High Precision |
| Q5_K_M | 5.69 | 95.27 GiB | 98.41 GiB | 103.66 GiB | Balanced High |
| Q4_K_M | 4.85 | 81.20 GiB | 84.35 GiB | 89.60 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 75.34 GiB | 78.49 GiB | 83.74 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 71.16 GiB | 74.30 GiB | 79.55 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 65.46 GiB | 68.61 GiB | 73.86 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 89.8 / 140.4 GiB | 112.2 tok/s | 0.50s | 64k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 89.6 / 199.7 GiB | 17.8 tok/s | 9.88s | 64k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 89.6 / 399.4 GiB | 17.8 tok/s | 9.88s | 64k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 89.6 / 149.8 GiB | 17.4 tok/s | 10.90s | 64k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 89.6 / 99.8 GiB | 11.7 tok/s | 15.03s | 64k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 89.6 / 99.8 GiB | 8.5 tok/s | 20.13s | 64k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | TIGHT | 89.8 / 95.0 GiB | 41.9 tok/s | 1.98s | 56k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | TIGHT | 89.6 / 96.0 GiB | 5.3 tok/s | 11.18s | 64k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | OOM | 89.8 / 79.4 GiB | OOM | — | OOM | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | OOM | 89.8 / 79.4 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5090 (32 GB) | OOM | 89.8 / 31.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 4090 (24 GB) | OOM | 89.8 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 89.8 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | OOM | 89.8 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 89.8 / 15.0 GiB | OOM | — | OOM | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | OOM | 89.8 / 23.0 GiB | OOM | — | OOM | Simulate → |
| Cloud NVIDIA L40S (48 GB) | OOM | 89.8 / 47.4 GiB | OOM | — | OOM | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | OOM | 89.8 / 46.6 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | OOM | 89.6 / 49.9 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 89.8 / 15.0 GiB | OOM | — | OOM | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | OOM | 89.6 / 49.9 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 89.8 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | OOM | 89.6 / 37.4 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | OOM | 89.6 / 49.9 GiB | OOM | — | OOM | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | OOM | 89.6 / 48.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 89.6 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 89.6 / 18.0 GiB | OOM | — | OOM | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | OOM | 89.6 / 48.0 GiB | OOM | — | OOM | Simulate → |
Interactive Mixtral 8x22B Instruct Simulator
141B (39B active) · GQA
819 GB/s · 58 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 64k
Memory Allocation Footprint 89.60 / 199.68 GiB usable
Weights: 81.20 GiB
KV Cache: 7.00 GiB
Overhead: 1.39 GiB
Headroom: 110.08 GiB
Decode Speed
17.8
tok/s (batch=1)
Prefill Speed
416
prompt tok/s
TTFT Latency
9.88s
for 4,096 tok
Active Params
39.0B
of 141B total
llama.cpp Server Command
llama-server -hf mistralai/Mixtral-8x22B-Instruct-v0.1:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.88 GiB | 83.05 GiB | 22.4 tok/s | COMFORTABLE |
| 8k | 1.75 GiB | 84.35 GiB | 21.6 tok/s | COMFORTABLE |
| 16k | 3.50 GiB | 86.10 GiB | 20.2 tok/s | COMFORTABLE |
| 32k | 7.00 GiB | 89.60 GiB | 17.8 tok/s | COMFORTABLE |
| 64k | 14.00 GiB | 96.60 GiB | 14.4 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
BUY LOCALOwn Hardware/mo
$113.43
267.8h active + power
Cloud GPU/mo
$1075.00
@$3.49/hr spot/pod
Serverless API/mo
$190.00
Payback: 20.5 mo
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Mixtral 8x22B Instruct
How much VRAM does Mixtral 8x22B Instruct require in Q4_K_M?
Mixtral 8x22B Instruct weights occupy 81.20 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 89.80 GiB.
What is the cheapest local hardware that runs Mixtral 8x22B Instruct?
Apple Mac Studio M3 Ultra (256 GB) (256 GB) runs Mixtral 8x22B Instruct Q4_K_M at ~17.8 tok/s decode and 9.88s TTFT.
How does Mixtral 8x22B Instruct scale KV cache at long context?
Mixtral 8x22B Instruct uses GQA with 8 KV heads across 56 layers (128 head_dim).